Whisper, LiveKit, Speechmatics, Gladia, Fish Audio and Vocode. All six can run on hardware you control, which is the only filter that matters when audio is not allowed to leave the building.
Five of the six also have a free tier on their hosted versions, so the decision is rarely about the licence fee. It is about who operates the thing.
What self-hosting replaces, and what it does not
A voice agent needs four components. Self-hosting can remove the per-minute line on some of them and cannot remove the phone call.
| Component | List rate | Self-hostable? |
|---|---|---|
| Speech recognition | $0.0077/min | Yes: Whisper, Speechmatics, Gladia |
| The model | ~$0.0108/min | Yes, with your own GPUs |
| Synthesis | ~$0.014/min | Yes: Fish Audio and others |
| Transport | $0.01/min hosted | Yes: LiveKit is open source |
| Telephony | $0.0140/min | No. A carrier bills this |
Telephony is the floor under the floor. Whatever you run yourself, a carrier still charges to place the call, and at Twilio's published $0.0140 that is about 30% of the $0.0465 component total.
The number that decides it is not on any pricing page
Operators running assembled stacks report figures well under the list-price floor. One published $0.01025 a minute on a stack using free-tier recognition and synthesis with a carrier and a model on top. Another reported $0.052 a minute across 17,284 minutes of real traffic, then gave a different figure in the same thread.
Treat all of those as reported claims rather than measurements. They are self-reported, usually by someone selling a service, and free tiers do not survive production volume. What they establish is a range, not a rate.
The cost self-hosting genuinely adds is the one nobody quotes: turn detection, barge-in, reconnection and model version pinning, all of which become yours. At small volume that engineering time exceeds any per-minute saving by a wide margin.
When it is the right answer
- Audio cannot leave your infrastructure. Healthcare, legal, government, or a customer contract that says so. This reason alone is sufficient and the arithmetic is irrelevant.
- Data residency in a specific jurisdiction that no hosted vendor offers for your region.
- Volume high enough that per-minute rates exceed the fully-loaded cost of running and maintaining the stack, computed with salary included.
- You already operate GPUs, so the marginal cost of another workload is close to zero.
Outside those four, hosted is usually cheaper once the engineering is priced honestly. The common mistake is comparing a licence fee of zero against a rate of $0.0077 and stopping there.